DAG-level `on_failure_callback` never fires

### Apache Airflow version 3.1.7 ### If "Other Airflow 3 version" selected, which one? _No response_ ### What happened? DAG-level `on_failure_callback` does not fire in Airflow 3.1.7 with KubernetesExecutor. Task-level `on_failure_callback` (via `default_args`) works correctly — it fires on the worker pod. ### Observations 1. **DagRun reaches `failed` state** — confirmed via `SELECT state FROM dag_run`. The `last_scheduling_decision` and `end_date` are within milliseconds of each other. 2. **`has_on_failure_callback` is present in serialized data** — confirmed via `SELECT data::text LIKE '%has_on_failure_callback%' FROM serialized_dag`. 3. **Scheduler DEBUG logs show `"callback is empty"` for every DagRun** — this is the log from `_send_dag_callbacks_to_processor` when `callback_to_run` is `None`, meaning `update_state()` returned no callback. 4. **DAG processor logs show no callback-related activity** — only routine DAG file parsing (`Setting next_dagrun ... to None`). ### What you think should happen instead? When `on_failure_callback` is set on a DAG and the DagRun fails, the callback should fire. This worked in Airflow 2.x. ### How to reproduce 1. Create a DAG with `on_failure_callback` set at the DAG level: ```python from datetime import datetime from airflow.sdk import DAG from airflow.providers.standard.operators.python import PythonOperator def my_callback(context): print("DAG failure callback fired") def task_that_fails(): raise ValueError("Intentional error") with DAG( dag_id="test-dag-level-callback", schedule=None, start_date=datetime(2025, 1, 1), catchup=False, on_failure_callback=[my_callback], ) as dag: PythonOperator( task_id="task_that_fails", python_callable=task_that_fails, ) ``` 2. Trigger the DAG. The task fails, the DagRun is marked as `failed`. 3. The DAG-level `on_failure_callback` never fires. No output in scheduler logs, DAG processor logs, or task logs. 4. Enable DEBUG logging on the sche…